Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence
Modeling Context Aware Dynamic Trust
Using Hidden Markov Model∗
Xin Liu
Anwitaman Datta
École Polytechnique Fédérale de Lausanne (EPFL)
1015 Lausanne, Switzerland
x.liu@epfl.ch
Nanyang Technological University
Singapore 639798
anwitaman@ntu.edu.sg
Abstract
to capture an agent’s dynamic behavior patterns, which is
the norm in large-scale open systems. For instance, in an
online auction site, a malicious seller may act honestly in
selling cheap items to gather sufficient reputation and then
cheat in selling an expensive item. Recently, several Hidden
Markov Model (HMM) based approaches (Moe, Tavakolifard, and Knapskog. 2008; Malik, Akbar, and Bouguettaya
2009; ElSalamouny, Sassone, and Nielsen 2009) have been
proposed to handle agents’ dynamic behaviors. These approaches model trust as a dynamic variable, changing with
time. An agent may switch from one state in the previous
interaction to another state in the next interaction. That is,
the trustor is able to deduce the trustworthiness state of the
target agent based on its past interactions. Since the system
state is hidden, HMM is naturally applied to infer the state
probability distribution which can be used to estimate outcome of the next interaction with the target agent.
Modeling trust in complex dynamic environments is
an important yet challenging issue since an intelligent
agent may strategically change its behavior to maximize its profits. In this paper, we propose a context
aware trust model to predict dynamic trust by using a
Hidden Markov Model (HMM) to model an agent’s interactions. Although HMMs have already been applied
in the past to model an agent’s dynamic behavior to
greatly improve the traditional static probabilistic trust
approaches, most HMM based trust models only focus
on outcomes of the past interactions without considering interaction context, which we believe, reflects immensely on the dynamic behavior or intent of an agent.
Interaction contextual information is comprehensively
studied and integrated into the model to more precisely
approximate an agent’s dynamic behavior. Evaluation
using real auction data and synthetic data demonstrates
the efficacy of our approach in comparison with previous state-of-the-art trust mechanisms.
Although HMM is potentially a good tool to model
dynamic trust, its accuracy greatly depends on observations probability distribution. Most existing approaches intuitively use the outcomes of the past transactions as the observation sequence. This method is effective when an agent
changes its behavior frequently or in specific patterns, but
is not well suited to identify relatively infrequent dishonest behavior of an agent who has a good behavior record,
or the implicit patterns from the random behaviors. For instance, when most of an agent’s past transactions are satisfactory, it is quite challenging to detect its ‘sudden’ behavior
change. In this paper, we first argue that by carefully investigating interaction contextual information, an agent’s dynamic behavior can be reflected better. We provide a comprehensive discussion on contextual information of an interaction using online auction site as the demonstration example. Such contextual information is characterized by a
set of features. In order to achieve efficient prediction, we
apply information theory (entropy based information gain)
and multiple discriminant analysis (MDA) to select the most
useful features and combine the same to generate a compact and effective feature vector, which is viewed as the
observations associated with each interaction. We then propose a HMM based trust model considering such contextual information to capture dynamic behavior of the target
agent. Specifically, we assume the outcome of an interaction has x levels (e.g., good, medium, bad). Each outcome
Introduction
In recent years, people have taken a more active role in
participating in various large-scale, open and dynamic systems like social networks, Peer-to-Peer (P2P) systems, ecommerce, etc. By intensively interacting with other system participants, user’s Internet experience has been significantly enhanced. However, due to the system characteristics
like openness, anyone can easily join the systems, it is thus
challenging to ensure a reliable interaction environment for
honest users. Traditional security solutions like public key
infrastructure (PKI) may help but they are not applicable to
address uncertainty of the user behavior in the interactions.
Trust management has emerged as a popular alternative to
reason about uncertainty, thus aiding decision support by indicating the trustworthiness of interaction partners.
A lot of approaches have been proposed to estimate trust
based on the target agent’s past behavior (Jøsang and Ismail
2002; Regan, Poupart, and Cohen 2006; Teacy et al. 2006;
Li and Wang 2010). However, most such approaches assume a relatively static agent behavior, and thus are not able
∗
The first author did most of this work at Nanyang Technological University, Singapore
c 2012, Association for the Advancement of Artificial
Copyright Intelligence (www.aaai.org). All rights reserved.
1938
level corresponds to a trustworthiness state of the target
agent. We thus construct the x-state HMM where the observable is the contextual information associated with each
interaction. Experimental results show that our approach is
more effective in detecting an agent’s dynamic behavior than
state-of-the-art trust approaches (Zhang and Cohen 2008;
Moe, Tavakolifard, and Knapskog. 2008; ElSalamouny, Sassone, and Nielsen 2009).
The rest of this paper is organized as follows: Section 2
discusses the related works. In Section 3, we provide a comprehensive study on trust contextual information using online auction site as the demonstration example. In Section
4, we propose a HMM based trust model. We first introduce the basic notations and background of HMM in Section 4.1 and then discuss the theory of our HMM based trust
model in Section 4.2. Section 4.3 elaborates how to integrate
contextual information into the model. Evaluation using real
auction dataset and synthetic dataset is conducted to quantify the performance of the proposed approach in Section 5.
Section 6 summarizes this work and discusses future work.
and (Malik, Akbar, and Bouguettaya 2009) demonstrate how
HMM based trust approaches are applied to distinct application scenarios: routing protocol design in mobile and ad-hoc
networks (MANET), and web service providers selection.
However, these works do not make full use of the contextual
information, which may also reflect the dynamic behavior of
an agent.
Different from HMM based approaches, Liu et al. (Liu
and Datta 2011) proposed to model an agent’s dynamic behavior by learning its past behavior patterns. Specifically,
the authors first identified features which are capable of describing context of an interaction. These features are then
used to calculate similarity between context of the two interactions. Trustworthiness of the potential transaction is estimated based on outcome of the specific past transaction
which has the most similar context. The main limitation of
this work is that it greatly relies on the sequence of the past
transactions. Our approach also makes use of the contextual information of the interactions (but in a different way),
which will be discussed in the next sections.
Related Work
Interaction Contextual Information
Modeling dynamic trustworthiness of an agent is important yet challenging (Castelfranchi 2011). Some early efforts on this issue is to extend the popular Beta (or Dirichlet) based trust models by adopting the “forgetting factor” (Jøsang and Ismail 2002; Buchegger and Boudec 2004;
Teacy et al. 2006). Briefly, these approaches pay more attention to the recent encounters than the old ones. This, to some
extent, reflects an agent’s recent behavior variation. In advance of “forgetting factor”, Zhang et al. (Zhang and Cohen
2008) took into account an agent’s dynamic behavior by introducing the concept of time window. The ratings of the target agent are partitioned into different elemental time windows. In each time window, the trustor counts the numbers
of successful and unsuccessful transactions. The trustworthiness of the target agent is firstly calculated by aggregating numbers of successful and unsuccessful transactions in
each time window (also taking into account forgetting rate)
and then is adjusted according to reputations of the indirect
experience providers.
Recently, several works applied HMM to model dynamic
trust. In (Moe, Tavakolifard, and Knapskog. 2008), a trust
model for multi-agent systems is developed to help the agent
make optimal trust decisions over time in a dynamic environment. The target agents’s behavior is predicted according to the HMM trust estimation module following the
Q-learning greedy policy. EISalamouny et al. (ElSalamouny, Sassone, and Nielsen 2009) modeled the real dynamic behavior of an agent by HMMs. They further justified the consistency of the model by measuring the difference between real and estimated predictive probability distributions using relative entropy. The work (Moe, Helvik,
and Knapskog 2009) conducted a comparison study between HMM based trust approaches and Beta based trust
approaches (with forgetting factor). The results show that
HMM based approach performs better in detecting agent’s
behavior changes thus is more realistic for dynamic environments. The works (Moe, Helvik, and Knapskog 2008)
From the perspective of behavior science and psychology
(Coons and Leibowitz 2010), we believe that an agent’s behavior change in the interactions is correlated with and can
be inferred (to certain extent) by the associated interaction
contextual information. For instance, in an online auction
site like eBay, a seller may vary his behavior consciously
or unwittingly in selling different items (e.g., he may be
very careful in selling very expensive items while be imprudent when trading individual cheap items). Next, using online auction site as the demonstration example, we provide a
comprehensive study on the interaction contextual information, which will be used in our HMM based trust model.
• About the target agent
The contextual information about the target agent includes the features collected from its profile in the system.
For instance, in online auction site, such features include
the seller’s system age, does he provide a detailed contact information, number of items already sold, reputation
value (e.g., fraction of successful transactions), average
delivery time, in which categories he is active and so on.
Besides its profile in the system, the context may also consist of the agent’s natural property including, for instance,
the seller’s age, gender, location (city/country), etc.
• About the provided services/products
The second class of contextual information is the features about the services (items in online auction site) provided by the target agent. Such features include the item
price, average item price in the same category, comments,
number of the same items in stock, number of comments
on these items, number of different buyers that already
placed bid on it, average age in the system of buyers that
already placed bid, etc.
• About the social relationships
The last class of contextual information relates to
the social relationships between the target agent and
1939
other agents. Examples of such relationships include family/colleague/friend/acquaintance ties, community/organization structure, trust networks (Jøsang, Hayward, and Pope 2006) and so on. For instance, the target
agent’s behavior may be inferred based on trustor’s experience with other similar agents (e.g., via stereotypes (Liu
et al. 2009)).
One basic problem that a HMM can solve is to find the
optimal state sequence for the underlying Markov process
given the past states/observations: P (xs |Xs , Ys , λ). With
this important property, we are able to predict the target
agent’s trustworthiness (i.e., outcome of the next interaction)
based on the past interactions and observations (i.e., the corresponding feature vectors).
We next present how to integrate the identified features
into the proposed HMM based trust model.
HMM Based Approach
Based on the properties of HMM, from the perspective
of the trustor ax , the target agent ay ’s dynamic trustworthiness can be modeled by a finite-state HMM. Such a
model is then used to estimate the predictive probability
distribution of ay ’s next state. Since we assume that outcome of a transaction is a discrete quantitative variable in
a certain range L, the l-state HMM constructed by ax for
modeling dynamic behavior of ay is denoted by λlx,y =
(L, Fx,y , π x,y , Ax,y , B x,y ).
Now we handle the problem of estimating outcome of the
next transaction with ay given sequence of outcomes of the
past transactions and the associated observed features. Here
the state of ay at discrete time point t corresponds to the rating of outcome Li ∈ L of the transaction θat x ,ay happened at
time t. Let Hm = s0 s1 s2 ...sm−1 (where ∀i=0,1,...,m−1 si ∈
L) be a random variable representing any sequence of outcomes of transactions between ax and ay where s0 is the outcome of the oldest transaction and sm−1 is the outcome of
the most recent one. We then T
denote the next transaction by
sm (and hence Hm+1 = Hm sm ). The corresponding observed features associated with each transaction
T is denoted
by Fm = f0 f1 f2 ...fm−1 , and Fm+1 = Fm fm (where
∀i=0,1,...,m fi ∈ F). The estimated probability distribution
of outcome of the next transaction given the HMM λlx,y is
thus obtained:
HMM Based Trust Model
The Basics
Consider a scenario where an agent ax , a service requestor
encounters a potential service provider ay . A transaction
happens when ax accepts ay ’s service. To indicate quality
of a service, ax may rate the transaction, where the rating is
a discrete quantitative variable in a certain range, denoted by
L = {L1 , L2 , ..., Ll }. For instance, the rating could be in the
range of [1, 2, 3, 4, 5], where 1 to 5 represents lowest quality, low quality, medium, high quality and highest quality
respectively. θax ,ay denotes the transaction between ax and
ay . We assume that ax maintains a list of past transactions1
with ay : Θax ,ay = {θa1x ,ay , θa2x ,ay , ...}.
As described in the introduction, we use HMMs to model
and approximate the dynamic behavior of an agent. A HMM
is a probabilistic model in which the observation sequence
is a probabilistic function of a finite set of hidden states.
Briefly, a (discrete) Hidden Markov Model (HMM) is defined as a tuple λ = (Q, V, π, A, B):
• Q = {q0 , q1 , ..., qN −1 } is the set of distinct states of the
Markov process, where N is the number of the states.
• V = {v0 , v1 , ..., vM −1 } is the set of M observation symbols.
• π is the initial state distribution.
P (sm = Lj |Fm+1 , λlx,y ) =
• A is the state transition probabilities (matrix): Q × Q →
[0,
P 1], with Ai,j = P (qj at time t + 1|qi at time t) and
qj ∈Q Ai,j = 1.
Since
• B is the observation probability matrix: Q × V →
[0,
P 1], with Bj,k = P (vk at time t|qj at time t) and
vk ∈V Bj,k = 1.
P (sm = Lj , Fm+1 , λlx,y )
.
P (λlx,y )
(3)
l
P
(F
,
λ
)
m+1
x,y
P (Fm+1 |λlx,y ) =
.
(4)
P (λlx,y )
P (sm = Lj , Fm+1 |λlx,y ) =
Consider a state sequence of length s: Xs =
(x0 , x1 , ..., xs−1 ), with the corresponding observations
Ys = (y0 , y2 , ..., ys−1 ). π0 is the initial probability stating
in state x0 . Then the probability distribution of the state sequence X given a HMM λ is obtained:
P (Xs |Ys , λ) = π0 Bx0 ,y0 Ax0 ,x1 Bx1 ,y1 ...Axs−2 ,xs−1
Bxs−1 ,ys−1 .
P (sm = Lj , Fm+1 , λlx,y )
.
P (Fm+1 , λlx,y )
(2)
We have
P (sm = Lj , Fm+1 |λlx,y )
.
P (Fm+1 |λlx,y )
(5)
P (sm = Lj , Fm+1 |λlx,y ) can be interpreted as the joint
probability that the sequence f0 f1 ...fm is observed and the
state sm of ay when the next transaction happens is Lj ∈
L; and P (Fm+1 |λlx,y ) is the probability of the observation
sequence Fm+1 , given the model.
By summing up all possible state sequences we have
P (sm = Lj |Fm+1 , λlx,y ) =
(1)
1
In case ax does not have sufficient past experience with ay , it
may resort to other agents who have interacted with ay . The issues
of addressing false feedback have been thoroughly studied (Teacy
et al. 2006; Zhang and Cohen 2008) thus is beyond the scope of
this work.
1940
Processing Contextual Information
P (Fm+1 |λlx,y ) =
X
P (Fm+1 , Hm+1 |λlx,y )
∀Hm+1
=
X
P (Fm+1 |Hm+1 , λlx,y )P (Hm+1 |λlx,y )
∀Hm+1
=
X
x,y
x,y
πsx,y
Bsx,y
Ax,y
s0 ,s1 Bs1 ,f1 · · · Asm−1 ,sm
0
0 ,f0
∀Hm+1
Bsx,y
.
m ,fm
(6)
However, such a computation is very expensive, thus is
infeasible in practice. In order to find P (Fm+1 |λlx,y ) (and
hence P (sm = Lj |Fm+1 , λlx,y )), we apply the forward algorithm (Rabiner 1989). For t = 0, 1, ..., m and k = 0, 1, ...,
l-1, we define
αt (k) = P (Ft+1 , st = Lk |λlx,y ).
Entropy(Θax ,ay ) = −
x,y
αt−1 (q)Ax,y
q,k ]Bk,t .
(7)
IGain(Θax ,ay , ωr ) = Entropy(Θax ,ay )−
X |Θυax ,ay |
Entropy(Θυax ,ay ).
|Θax ,ay |
P (Fm+1 , sm = Lk |λlx,y )
k=0
=
l−1
X
(9)
αm (k).
k=0
Finally, according to Eq. 5 and 9 we have the probability
distribution of the outcome of the next transaction with ay :
αm (j)
P (sm = Lj |Fm+1 , λlx,y ) = Pl−1
.
k=0 αm (k)
Pl−1
where j=0 P (sm = Lj |Fm+1 , λlx,y ) = 1.
We can then obtain the most likely outcome by:
sm = arg max[P (sm = Lj |Fm+1 , λlx,y )].
(13)
The information gain of a feature measures expected reduction in entropy by considering this feature. Clearly, the
higher the information gain, the lower the corresponding entropy becomes and thus the better the classification of past
transactions is achieved. Then we may select the top-K features that have the highest information gain3 .
After selecting the features, we apply multiple discriminant analysis (MDA) (McLachlan 2004) to combine the features to generate a smaller but more effective feature set. The
aim of the MDA is to find a transformation Φ that can maximize the inter class variance and minimize the intra class
variance4 . The original feature set Ω is then transformed as
Step 3: From definition of αt (k) (Eq. 7), it is straightforward that P (Fm+1 |λlx,y ) can be obtained by
l−1
X
(12)
υ∈Υ(ωr )
(8)
q=0
P (Fm+1 |λlx,y ) =
pj log2 pj .
Entropy is used to characterize (im)purity of a collection of examples. For each feature ωr ∈ Ω, we assume it
has a set of values (e.g., discrete variable) or intervals (e.g.,
continuous variable), which is denoted by Υ(ωr ). For each
υ ∈ Υ(ωr ), we denote the set of past transactions that are
associated with υ for feature ωr by Θυax ,ay . The information
gain of feature ωr is thus calculated by:
Step 2: For t = 0, 1, ..., m and k = 0, 1, ..., l-1, we derive
αt (k) = [
l−1
X
j=0
This is the joint probability of the observation subsequence
(from the first observation to the tth one), and the corresponding state st is Lk . The αt (k) is derived recursively:
x,y
Step 1: Let α0 (k) = πkx,y Bk,0
, for k = 0, 1, ..., l-1.
l−1
X
According to the discussion on trust contextual information
(see interaction contextual information section), we select
a set of features Ω = {ω1 , ω2 , ...} to describe/characterize
each transaction2 . These features are expected to have the
potential to distinguish different levels (∈ L) of the transaction outcome. To do such a feature selection, we use entropy based information gain (MacKay 2003). Given the past
transactions Θax ,ay = {θa1x ,ay , θa2x ,ay , ...} between ax and
ay , we denote the fraction of transactions with outcome Lj
by pj . Then the entropy of all past transactions Θax ,ay is:
0
Ω = ΦT Ω.
(14)
ay θai x ,ay
∈
In this way, each transaction between ax and
Θax ,ay is described/characterized by the transformed feature
0
set, denoted by Ωax ,ay ,i . Such a transformed feature set is
actually the observation in our HMM based trust model.
Finally, the observation probability matrix B x,y is derived
based on the outcomes of the past transactions and the associated transformed feature sets. Note that in case the feature
values are continuous, we resort to the methods (e.g., C4.5
(Quinlan 1993)) used in decision tree to handle the continuous feature.
(10)
(11)
Lj ∈L
To conduct the computation above, we need to learn the
state transition matrix Ax,y and the observation probability matrix B x,y . Ax,y can be easily derived based on the
outcomes of the previous transactions. Different from traditional HMM based approaches which treat transaction outcome as the observation, our approach uses transaction contextual information associated with each transaction as the
observations. We next detail how to integrate such contextual information into our HMM based trust model.
2
Feature selection is application dependent. We will show what
features are selected using online auction site as the example in the
next section.
3
Alternative way may choose features by setting a information
gain threshold.
4
Each class represents a set of transactions with a certain level
of outcome Łj ∈ L
1941
Evaluation
Results
We first compare our model with traditional HMM based
trust model (e.g., (ElSalamouny, Sassone, and Nielsen
2009)) using Allegro dataset. We study how false positive
rate and false negative rate evolve with different volumes of
HMM training data. That is, for each (of 150) selected target
seller, we use first x% of the previous transactions to build
a HMM (initially x = 50) by which we predict the outcome of the next transaction. From Fig. 1 we observe that
for both approaches, false positive rate and false negative
rate decrease with the increasing fraction of transactions for
HMM construction. This is mainly because with more training data, HMM is able to more accurately estimate probability distribution over all possible dynamic behavior patterns,
and hence provide more precise prediction on the outcome
of the next transaction. We also observe that our trust model
incurs lower rate for both false positive and false negative
than traditional HMM based trust model. This validates that
by considering transaction contextual information (i.e., the
selected features) which correlates with the seller’s behavior
or intent, the trustor can identify certain implicit behavior
patterns that may be difficult to be detected by simply investigating outcomes of the previous transactions.
Experimental Methodology
We use real dataset collected from an Internet auction site
Allegro (http://allegro.pl/) as well as synthetic data to evaluate performance of our HMM based trust model. The Allegro dataset contains 10,000 sellers, 10,000 buyers, more
than 200,000 transactions and over 1.7million comments. In
order to fully understand how a seller changes its behavior in
the transactions, we select a set of (150) sellers which have
sufficient historical information (i.e., over 100 past transactions).
We assume binary outcome of a transaction, i.e., a transaction is considered to be successful if its feedback is positive, otherwise, it is considered to be unsuccessful. So we
construct a 2-state HMM. The features we select to characterize the context of each transaction include: (1) category of
the item, (2) difference between item price and average price
over the items in the same category, (3) number of items already sold by the seller at the time the transaction happened
and (4) reputation (i.e., fraction of successful transactions)
when the transaction happened. Note that we do not use the
first feature directly in the computation, but use it to collect
the relevant transactions to construct the model.
Real Allegro data provides a realistic environment, however, behavior patterns of the sellers in real data are predetermined (i.e., fixed), and for which we do not have ground
truth. In order to comprehensively evaluate performance
of our approach under different circumstances, and also to
more flexibly control agents’ behavior, we generate synthetic data derived from real data. Specifically, we generate
a synthetic seller with 100 past transactions. All transactions
with their outcomes and features are taken from real Allegro
dataset (i.e., a good/bad synthetic transaction is generated
from a randomly selected good/bad real transaction). We
simulate three types of dynamic behavior following the configurations of (Keung 2011) and (Moe, Helvik, and Knapskog 2009): (1) the agent behaves honestly in several transactions and then cheats in the next transaction; (2) the agent
changes its behavior half way (from honesty to dishonesty);
(3) the agent changes its behavior randomly (but still acting
honestly with a higher probability, say 70%).
We compare our approach with traditional HMM based
trust models (i.e., without contextual information)5 . The
metrics we use to evaluate performance of the approaches
include:
(a) False positive.
• false positive rate
The transaction is unsuccessful but the algorithm predicted that it would be successful.
(b) False negative.
• false negative rate
The transaction is successful but the algorithm predicted
that it would be risky.
Figure 1: Experiments using Allegro dataset.
We then conduct a comparison study of the HMM based
trust models under three types of simulated dynamic behaviors. Fig. 2(a) demonstrates performance of the approaches
when the target agent behaves honestly in several transactions (e.g., random value from 8 to 12) and then cheats in
5
Since the work (Moe, Helvik, and Knapskog 2009) has proven
that HMM based approaches are more effective than the ‘forgetting
factor’ based approaches, we thus only compare our approach with
existing HMM based trust models
1942
transactions6 (see Fig. 2(b)). However, our approach can
quickly learn the context of the transaction thus is able to
identify the unsuccessful transactions earlier than traditional
HMM based trust model.
From Fig. 2(c) we observe that when the target agent behaves randomly (i.e., behavior patter 3), our model still outperforms traditional HMM based approach. The reason for
this is that although traditional HMM based approach is capable of detecting explicit behavior patterns, it is difficult for
this approach to identify implicit patterns from the random
behaviors. In contrast, our trust model improves traditional
approach by considering transaction contextual information
that correlates with the agent’s behavior. Another interesting
phenomenon is that different from the other results (see Fig.
2(a) and 2(b)) where more training data (i.e., past transactions) evidently lowers the falseness rates, in this scenario,
the improved accuracy resulting from larger volume of training data is limited. This is because even if more past transactions are used to build a HMM, it is not easy to detect
implicit patterns from the random behaviors thus restricting
the improvement on the prediction accuracy (for both approaches).
the next transaction. Similar to the results using real Allegro data, for both approaches, false positive rate and false
negative rate decrease as training data (for HMM) volume
increases. It is obvious that our model is more accurate than
traditional HMM based trust model. This is mainly because
most of the past transactions are successful, predicting outcome of the next transaction simply based on outcomes of
the previous transactions is likely to ‘miss’ a potentially
risky transaction.
(a) Dynamic behavior pattern (1).
Conclusion
This paper presented a HMM based context aware trust
model. The dynamic behavior of an agent is approximated
by a finite state HMM. Different from traditional HMM
based trust approaches that rely on a sequence of outcomes
of the past interactions with the target agent to estimate probability distribution over the outcome of the next interaction,
our approach investigates and utilizes the interaction contextual information (as the observation) to build a HMM. Information theory (i.e., information gain) and machine learning technique (i.e., multiple discriminant analysis) are applied to select and process the contextual information to
achieve accurate prediction. This strategy can help in revealing the behavior patterns which cannot be identified by simply statistically studying the outcomes of the previous interactions. Real auction dataset and synthetic data based evaluation demonstrates that our approach is more effective in
detecting various dynamic behavior patterns than traditional
HMM based approaches that do not consider contextual information.
In the future, we aim to apply the proposed trust model
to various multi-agent systems to validate its practicability.
Specifically, analyzing the complexity and studying the optimization issues according to different application scenarios
would be the next research focuses.
(b) Dynamic behavior pattern (2).
(c) Dynamic behavior pattern (3).
Acknowledgement
Figure 2: Synthetic dataset.
The authors are grateful to Dr. Adam Wierzbicki of Polish
Japanese Institute of Information Technology for his assistance in providing Allegro dataset.
When the second type of behavior pattern (i.e., seller acts
honestly for the first half transactions and acts dishonestly
for the rest ones) is applied, we observe that neither approaches can detect dishonest behavior for the first several
6
Note that since we only predict the second half transactions so
only false positive rate is shown.
1943
References
Rabiner, L. R. 1989. A tutorial on hidden markov models
and selected applications in speech recognition. In Proceedings of the IEEE, 257–286.
Regan, K.; Poupart, P.; and Cohen, R. 2006. Bayesian reputation modeling in e-marketplaces sensitive to subjecthity,
deception and change. In Proceedings of the 21st national
conference on Artificial intelligence - Volume 2.
Teacy, W. T. L.; Patel, J.; Jennings, N. R.; and Luck, M.
2006. Travos: Trust and reputation in the context of inaccurate information sources. Autonomous Agents and MultiAgent Systems 12(2):183–198.
Zhang, J., and Cohen, R. 2008. Evaluating trustworthiness
of advice about seller agents in e-marketplaces: A personalized approach. Electronic Commerce Research and Applications 7(3):330–340.
Buchegger, S., and Boudec, J.-Y. L. 2004. A robust reputation system for peer-to-peer and mobile ad-hoc networks.
In Proceedings of the Second Workshop on the Economics
of Peer-to-Peer Systems.
Castelfranchi, C. 2011. Trust: nature and dynamics. In Proceedings of the 9th ACM SIGCHI Italian Chapter International Conference on Computer-Human Interaction: Facing
Complexity, CHItaly.
Coons, E. E., and Leibowitz, S. F. 2010. Neal e. miller and
his research. Biofeedback 38(3):101–107.
ElSalamouny, E.; Sassone, V.; and Nielsen, M. 2009. Hmmbased trust model. In Formal Aspects in Security and
Trust’09, 21–35.
Jøsang, A., and Ismail, R. 2002. The beta reputation system. In Proceedings of the 15th Bled Electronic Commerce
Conference.
Jøsang, A.; Hayward, R.; and Pope, S. 2006. Trust network
analysis with subjective logic. In Proceedings of the 29th
Australasian Computer Science Conference - Volume 48.
Keung, S. N. L. C. 2011. Trust-based social mechanism to
counter deceptive behaviour. Ph.D. Dissertation, University
of Warwick.
Li, L., and Wang, Y. 2010. Subjective trust inference in
composite services. In Proceedings of the 24th AAAI.
Liu, X., and Datta, A. 2011. A trust prediction approach
capturing agents’ dynamic behavior. In Proceedings of the
22nd IJCAI.
Liu, X.; Datta, A.; Rzadca, K.; and Lim, E. P. 2009.
Stereotrust: a group based personalized trust model. In Proceeding of the 18th CIKM.
MacKay, D. J. C. 2003. Information theory, inference, and
learning algorithms. Cambridge University Press.
Malik, Z.; Akbar, I.; and Bouguettaya, A. 2009. Web services reputation assessment using a hidden markov model.
In Proceedings of the 7th International Joint Conference on
Service-Oriented Computing, ICSOC-ServiceWave ’09.
McLachlan, G. J. 2004. Discriminant Analysis and Statistical Pattern Recognition. Wiley-Interscience.
Moe, M. E.; Helvik, B. E.; and Knapskog, S. J. 2008. Tsr:
trust-based secure manet routing using hmms. In Proceedings of the 4th ACM symposium on QoS and security for
wireless and mobile networks.
Moe, M. E. G.; Helvik, B. E.; and Knapskog, S. J. 2009.
Comparison of the Beta and the Hidden Markov Models of
Trust in Dynamic Environments. In Proceedings of the 3rd
IFIP WG11.11 International Conference on Trust Management.
Moe, M. E. G.; Tavakolifard, M.; and Knapskog., S. J. 2008.
Learning trust in dynamic multiagent environments using
hmms. In Proceedings of The 13th Nordic Workshop on Secure IT Systems (NordSec2008).
Quinlan, J. R. 1993. C4.5: programs for machine learning. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.
1944